Pith. sign in

REVIEW 4 major objections 6 minor 34 references

Robust Relevance Feedback for Interactive Known-Item Video Search

T0 review · 4 major / 6 minor · reviewed 2026-08-07 · deepseek-v4-flash

Challenge this review Re-run · record.json Download PDF Read on arXiv ↗ Review JSON

Pith's one-line read A system that predicts which of several similarity perceptions a user is relying on can promote most hard-to-find videos to the top rank in seven rounds of pairwise feedback.

desk verdict A competent KIS relevance feedback paper with a genuinely novel combination, but the robustness claim is untested because the simulated user is trained and evaluated on the same majority-vote labels. read the letter →

arxiv 2505.15128 v1 pith:J2RFEBJJ submitted 2025-05-21 cs.IR cs.MM

Zhixin Ma , Chong-Wah Ngo This is my paper · ORCID
classification cs.IRcs.MM
keywords Known-itemsearchrelevancefeedbackvideoretrievalpairwisejudgmentuserperceptionmodelingBayesianupdateinteractiveV3Cdataset
verification ladder T0 review T1 audit T2 compute T3 formal

The pith

A machine-rendered reading of the paper's core claim, the machinery that carries it, and where it could break.

The reading

The paper claims that interactive known-item video search can be made robust to inconsistent user feedback by replacing top-k selections with pairwise relative judgments and by treating user perception as a mixture of several embedding-based sub-perceptions. A predictive model assigns a confidence score to each sub-perception for each feedback instance, and the Bayesian update uses those scores to down-weight judgments that are misaligned with the user's apparent choice. On the million-scale V3C collections, the method promotes more than 60% of targets with initial ranks 10-50 to the top rank within seven rounds, and more than 40% of targets initially ranked 1,000-5,000. With an explicit pruning of the initial search space to the top 5,000 candidates, overall Recall@1 reaches 0.6384. If the result holds, it would mean relevance feedback remains useful in single-target search even when users cannot reliably mirror the machine's similarity measure.

What carries the argument

The load-bearing object is the pairwise relative judgment expressed as a feature-difference vector, $v_{\text{diff}} = v_{+} - v_{-}$, where $v_{+}$ and $v_{-}$ are the representations of the chosen and unchosen videos in a pair. The paper observes that the similarity $s(v_{\text{diff}}, v_i)$ is proportional to the term $\left[s(v_{+}, v_i) - s(v_{-}, v_i)\right]$ that drives PicHunter's Bayesian update, so the difference vector is a natural input for predicting which sub-perception the user is following. A transformer-based predictive user model consumes the query, the interaction history, the search state, and a learnable distance embedding of $\|v_{\text{diff}}\|$ (quantized into 100 bins) to output per-sub-perception confidence scores, and a soft Bayesian update multiplies each sub-perception's evidence by the confidence score. This combination is what lets the system down-weight misaligned sub-perceptions instead of assuming the user always agrees with the machine.

What would settle it

Have human users make the same pairwise relative judgments on a sample of displayed pairs from V3C2, and check whether the model's per-sub-perception confidence scores are calibrated to which embedding feature agrees with the human choice; if human judgments do not track the majority-vote simulator or the confidence scores do not predict agreement, the reported robustness collapses.

Watch

Extended reading notes

Core claim

Known-item search has exactly one target, so standard relevance feedback's positive examples are unavailable; the paper's central claim is that the classic PicHunter Bayesian update can still be made robust by letting the user give pairwise relative judgments ('which of these two is closer to the target?') and by modeling the user's perception as a weighted combination of sub-perceptions, each living in a separate embedding space (CLIP4Clip, ITV, and BLIP). A transformer-based predictor consumes the initial query, the current search state, and the interaction history to output a confidence score for each sub-perception at each feedback, and the Bayesian update weights each sub-perception's similarity evidence by that confidence, filtering out sub-perceptions that disagree with the user's choice. The paper reports that this raises seven-step Recall@1 on V3C2 to 0.5467 from PicHunter's 0.4949, that over 60% of targets initially ranked 10-50 and over 40% of targets initially ranked 1,000-5,000 reach rank 1, and that pruning candidates below rank 5,000 lifts Recall@1 to 0.6384.

Load-bearing premise

The evaluation and the training signal both assume that a user's inconsistent judgment is the majority vote of the three embedding features that define the sub-perceptions; if real human disagreement does not follow that rule, the claimed robustness is not tested.

Editorial extensions

If this is right

  • Within seven rounds of pairwise feedback, the model moves over 60% of targets with initial rank 10-50 and over 40% of targets with initial rank 1,000-5,000 to the top position.
  • Overall Recall@1 after seven steps is 0.5467, above PicHunter's 0.4949; with search-space pruning to the top 5,000 initial candidates it reaches 0.6384.
  • On the 17 VBS textual KIS queries, the model ranks 16 targets at top-1 with the full three-round query text and all 17 targets within top-10.
  • The choice of captioning model (BLIP2 vs LLaVA-NeXT-Video) has minimal impact on Recall@1 across all search-depth ranges.
  • Ablations show that removing the soft weighted update hurts most (0.5467 to 0.3321), while removing the distance embedding costs less but the gap grows over rounds.

Reading between the lines

Editorial extensions of the paper, not claims the author makes directly.

  • We infer that the same perception-filtering idea should transfer to ad-hoc video search with multiple relevant items, where a user's notion of relevance may also switch between sub-perceptions across a session.
  • We infer that the reported 'rank 1' after pruning is rank among the pruned candidates, so the gap between pruned and unpruned success suggests some targets below rank 5,000 are simply excluded rather than retrieved.
  • We infer a testable scaling property: the method's advantage over PicHunter should grow when the three embedding features disagree on displayed pairs and shrink when they agree, because the predictive model only has something to filter when sub-perceptions conflict.
  • We infer that the pairwise difference representation $v_{\text{diff}} = v_{+} - v_{-}$ could be reused as a generic preference signal in other interactive ranking systems, independent of the Bayesian update it feeds here.
Share X Bluesky LinkedIn Reddit HN

Signed reviews

No signed human review yet.

Editorial analysis

A structured set of objections, weighed in public.

Desk editor's note, referee report, and a circularity audit.

Referee Report

4 major / 6 minor

Summary. The paper addresses known-item video search with relevance feedback. It replaces PicHunter's top-k selection with pairwise relative judgments, decomposes user perception into three sub-perceptions (CLIP4Clip, ITV, and BLIP embeddings), and trains a predictive user model to assign confidence weights to these sub-perceptions before a soft Bayesian update. The system is evaluated on V3C1/V3C2 with a simulated user that selects the majority-vote choice among the three embedding-based distance comparisons. The paper reports Recall@1 improvements over PicHunter, including 0.5467 without pruning and 0.6384 with pruning after seven steps, and claims that targets initially ranked between 1,000 and 5,000 are promoted to rank 1 with success rate exceeding 40%.

Significance. If the claims were supported, the idea of learning to weight multiple embedding subspaces from user feedback would be a useful contribution to interactive video retrieval. The paper also provides a large-scale evaluation on V3C, an ablation study, and results on VBS t-KIS queries. However, the central claim of robustness to inconsistent human feedback is not established by the experiments: the simulated user that generates feedback at evaluation time is the same majority-vote rule that defines the training labels for the predictive model, so the reported gains may only reflect the model's ability to fit this synthetic rule. The internal contradiction between the abstract's 'success rate exceeding 40%' and the conclusion's '6.4%' for the same depth range further undermines the headline results. No real-user study is reported, despite the paper's emphasis on human inconsistency.

major comments (4)
  1. [§4.3, §4.5, §5.3] The evaluation is circular. Section 4.3 defines the user simulator as the majority vote of three sub-perceptions (CLIP4Clip, ITV, BLIP), and Section 4.5 trains the predictive user model with BCE labels y_i that indicate whether each feedback agrees with that same majority vote. Section 5.3 then evaluates the full system with the identical simulator. Consequently, the test-time feedback is drawn from exactly the distribution the model was optimized to predict. The reported Recall@1 gains over PicHunter therefore demonstrate that the model can learn to mimic the synthetic majority-vote rule, not that it is robust to real human inconsistency. The abstract's claim that the method works 'despite inconsistent feedback' is unsupported. A real-user study, or at minimum an independent noise model that is not derived from the training labels, is needed to substantiate the robustness claim.
  2. [Abstract vs. §6 (Conclusion)] The headline numbers are internally inconsistent. The abstract states that for targets initially ranked between 1,000 and 5,000, the model 'achieves a success rate exceeding 40%' in optimizing ranks to the top. The conclusion, however, says that 'at a depth beyond 1,000, our approach can boost 6.4% of search targets to top-1.' These cannot both be accurate descriptions of the same experimental setup unless 'success rate' and 'boost' refer to different quantities, which the paper does not explain. The introduction (Section 1) also says 'our method achieves a 40% chance of reaching rank-1' for the same depth range. The authors must correct this contradiction and state precisely which metric is reported for which depth interval.
  3. [§4.4.3, §5.4, Table 2] Search space pruning changes the meaning of the reported Recall@1 values. Section 4.4.3 proposes discarding all candidates outside the top 5,000 of the initial query ranking, and Section 5.4 applies this pruning before computing Recall@1. Thus 'rank 1' in the pruned results is relative to a 5,000-item subset, not to the full V3C2 collection, making the 0.6384 figure not directly comparable to the 0.5467 unpruned result. In addition, the evaluation queries are restricted to targets initially ranked no deeper than 5,000, so the method is not tested on targets below that depth even though the pruning step would exclude them from the search space entirely. The paper should report both unpruned and pruned results with the rank semantics made explicit, and discuss the coverage limitation.
  4. [§4.5, Eq. (7)] The description of model training is not sufficiently precise to establish what the predictor learns. The text says training uses 'simulated interactions in the perfect user setting' and then says the model is 'trained to predict the sub-perception of every user feedback at iteration t in the imperfect user setting,' but it does not define the label y_i in Eq. (7) beyond 'whether a feedback aligns with the sub-perception.' Given that Section 4.3 defines misalignment as disagreement with the majority vote, the y_i labels appear to be generated by the same rule as the test-time simulator. This should be stated explicitly, and the term 'perfect user setting' should be clarified, because it seems to contradict the construction of an imperfect user in Section 4.3.
minor comments (6)
  1. [§2] There is a typo in Section 2: 'struggles to to maintain its effectiveness' should read 'struggles to maintain its effectiveness.'
  2. [§4.4.1, Eq. (1), Eq. (6)] The notation 'EX P' appears in Eqs. (1) and (6); it should be formatted as 'exp' for readability and consistency with Eq. (4).
  3. [§4.4.3, §5.1] The pruning threshold is denoted N_prune in Section 4.4.3 but N_tprune in Section 5.1; please unify the notation.
  4. [§5.3, Table 1] Table 1 reports aggregate Recall@1 over all depths, while Figure 6 reports per-depth results. The text should clarify that the 0.5467 figure is an average across the five depth intervals, because readers may otherwise interpret it as a uniform success rate. The same applies to the pruning results in Table 2.
  5. [§5.5, Table 3] The formatting of Table 3, with 'Recall@1' and 'Recall@10' over three rounds, is hard to read because the column headers are not clearly separated; a layout with explicit per-round columns would improve clarity.
  6. [§5.3] The paired t-test is mentioned but no standard deviations or per-depth significance values are reported; providing these would strengthen the comparison with PicHunter.

Circularity Check

1 steps flagged · score 5.0 of 10

Central 'robustness' evaluation is circular: the predictive user model is trained and tested on the same majority-vote simulator it is built to fit; real-user validation is absent.

  1. fitted input called prediction [Introduction; Section 4.3 User Simulator; Section 4.5 Eq. (7); Section 5.5 VBS evaluation]
    "During training and evaluation, after the display is visualized, a user simulator provides relative judgments based on majority voting across sub-perceptions. With access to the search target, the user simulator can determine Oracle judgment by comparing their distances to the target within each sub-perception. The ultimate user judgment is simulated by selecting the majority choice among all sub-perceptions. The predictive user model needs to filter out the sub-perception(s) that are misaligned with the simulated judgments. We use binary cross entropy (BCE) as the loss function..."

    The labels y_i in Eq. (7) are defined as agreement with the same majority-vote over the three sub-perception embeddings (CLIP4Clip, ITV, BLIP) that also generates every test-time user judgment (Sec. 5.5: 'The user simulator applies majority voting based on the similarity measure of three sub-perceptions'). Thus the predictive user model is fitted to the exact rule that produces both its supervision and its evaluation feedback. The reported Recall@1 improvements and the robustness claim 'despite inconsistent feedback' are in-distribution results against a self-generated user; they do not provide independent evidence about how the system behaves with real human inconsistency. The 'prediction' is a fit to the simulator, not a prediction validated outside the system's own definition.

full rationale

The derivation of the pairwise feedback and soft Bayesian update is self-contained: Eqs. (1)-(6) are algebraic consequences of PicHunter's update with confidence weighting, and the distance-embedding argument is a rearrangement of the cosine-similarity numerator. No uniqueness theorem is imported from the authors' prior work, and the citations to the authors' own VBS papers are not load-bearing. The circularity is confined to the evaluation of the robustness claim. Both the supervision for the predictive user model (Eq. (7), y_i = 'whether a feedback aligns with the sub-perception') and the test-time user are generated by the same majority vote over the three sub-perception embeddings; hence the model is fitted to the rule that defines its own evaluation. This makes the headline 'robustness despite inconsistent feedback' an in-distribution fit rather than a test against independent human judgments, warranting a partial-circularity score. Separately, the abstract's 'exceeding 40%' for targets initially ranked 1,000-5,000 conflicts with the conclusion's 6.4% at depth beyond 1,000, and the search-space-pruning numbers report rank-1 within a 5,000-item retained set; these are correctness and disclosure concerns rather than circularity.

Assumptions & free parameters 7 free parameters · 5 assumptions · 1 invented entities

The central claim depends on several hand-chosen hyperparameters (rho, N_prune, N_D, display size, distance embedding granularity) and on domain assumptions about how users and the simulated user behave. The most consequential item is the assumption that the majority vote of three fixed embeddings is an adequate stand-in for human perceptual inconsistency, since both the training labels for the predictive model and the evaluation feedback come from that same simulator. The pruning threshold N_prune materially changes the evaluation protocol, because rank-1 after pruning is rank-1 within a top-5,000 shortlist, not necessarily within the full collection. The sub-perception decomposition is an invented modeling construct without independent evidence.

free parameters (7)
  • rho = 0.05
    Sharpness parameter in the Bayesian update (Eq. 6); chosen by hand, not learned.
  • N_prune = 5,000
    Search space pruning threshold (Sections 4.4.3 and 5.4); major effect on results (0.5467 to 0.6384), selected ad hoc.
  • N_D = 100
    Candidate pool size for diverse display (Section 4.2); chosen by the authors.
  • display_pairs = 5
    Number of video pairs shown per iteration (Section 5.1); fixed by design.
  • distance_embedding_intervals = 100
    Number of intervals for the L2-norm distance embedding (Section 4.4.1); chosen by hand.
  • state_top_videos = 50
    Number of top-ranked videos pooled to represent the search state (Section 4.4.1); chosen empirically.
  • max_iterations = 7
    Number of feedback rounds in training and evaluation (Section 5.1); set to limit interaction cost.
assumptions (5)
  • domain assumption The simulated user's relative judgment is the majority vote over the three sub-perception embeddings.
    Section 4.3 defines the user simulator this way; all training and evaluation feedback is generated by this rule.
  • ad hoc to paper User perception decomposes into sub-perceptions aligned with CLIP4Clip, ITV, and BLIP feature spaces.
    Sections 3 and 4.3 assume the user's intent matches one or several of these fixed embeddings; no human evidence is provided.
  • ad hoc to paper Candidates outside the top 5,000 of the initial query ranking can be safely discarded.
    Section 4.4.3 justifies this as engineering to stabilize the search state; it changes the rank-1 criterion in pruned experiments.
  • domain assumption PicHunter's exponential Bayesian update (Eq. 1) is a valid way to refine the target distribution.
    The paper adopts the update rule from PicHunter [2-4] without re-deriving it.
  • domain assumption Cosine similarity in each embedding space reflects perceptual closeness to the target.
    Used in Eqs. (1)-(6) for both the simulator and the update; not validated against human similarity judgments.
invented entities (1)
  • Sub-perception decomposition of user preference
    purpose: Represents the user's perceptual state as a weighted combination of independent embedding spaces, enabling the model to filter out misaligned sub-perceptions.
    The paper postulates that users align with one or several of the three chosen embedding spaces (Sections 3 and 4.3) without external evidence; it is a modeling construct that drives the whole pipeline.

how reviews work

0 comments
Cite this review

Pith. "Pith review of Robust Relevance Feedback for Interactive Known-Item Video Search." pith.science (2026). https://pith.science/paper/J2RFEBJJ

@misc{pith2026250515128,
  author       = {Pith},
  title        = {Pith review of: Robust Relevance Feedback for Interactive Known-Item Video Search},
  year         = {2026},
  howpublished = {\url{https://pith.science/paper/J2RFEBJJ}},
  note         = {Machine review of arXiv:2505.15128}
}
read the original abstract

Known-item search (KIS) involves only a single search target, making relevance feedback-typically a powerful technique for efficiently identifying multiple positive examples to infer user intent-inapplicable. PicHunter addresses this issue by asking users to select the top-k most similar examples to the unique search target from a displayed set. Under ideal conditions, when the user's perception aligns closely with the machine's perception of similarity, consistent and precise judgments can elevate the target to the top position within a few iterations. However, in practical scenarios, expecting users to provide consistent judgments is often unrealistic, especially when the underlying embedding features used for similarity measurements lack interpretability. To enhance robustness, we first introduce a pairwise relative judgment feedback that improves the stability of top-k selections by mitigating the impact of misaligned feedback. Then, we decompose user perception into multiple sub-perceptions, each represented as an independent embedding space. This approach assumes that users may not consistently align with a single representation but are more likely to align with one or several among multiple representations. We develop a predictive user model that estimates the combination of sub-perceptions based on each user feedback instance. The predictive user model is then trained to filter out the misaligned sub-perceptions. Experimental evaluations on the large-scale open-domain dataset V3C indicate that the proposed model can optimize over 60% search targets to the top rank when their initial ranks at the search depth between 10 and 50. Even for targets initially ranked between 1,000 and 5,000, the model achieves a success rate exceeding 40% in optimizing ranks to the top, demonstrating the enhanced robustness of relevance feedback in KIS despite inconsistent feedback.

Figures

Figures reproduced from arXiv: 2505.15128 by the authors.

Figure 1
Figure 1. Framework of the proposed relevance feedback system. The initial query for the search target is “a woman in a red [PITH_FULL_IMAGE:figures/full_fig_p004_1.png] view at source ↗
Figure 2
Figure 2. Architecture of the proposed predictive model [PITH_FULL_IMAGE:figures/full_fig_p005_2.png] view at source ↗
Figure 3
Figure 3. Illustration of the shift in search results. The initial [PITH_FULL_IMAGE:figures/full_fig_p006_3.png] view at source ↗
Figures from the paper (2 more)
Figure 4
Figure 4. Figure 4: Example of LLaVA-NeXT and BLIP2 captions. [PITH_FULL_IMAGE:figures/full_fig_p007_4.png]
Figure 5
Figure 5. Figure 5: Performance comparison of different caption [PITH_FULL_IMAGE:figures/full_fig_p007_5.png]

Discussion (0). Continue with ORCID to comment.

Reference graph

Works this paper leans on

34 extracted references · 26 canonical work pages

  1. [1]

    Fabian Berns, Luca Rossetto, Klaus Schoeffmann, Christian Beecks, and George Awad. 2019. V3C1 Dataset: An Evaluation of Content Characteristics. In Proceed- ings of the 2019 on International Conference on Multimedia Retrieval (Ottawa ON, Canada) (ICMR ’19). Association for Computing Machinery, New York, NY, USA, 334–338. doi:10.1145/3323873.3325051

    work page arXiv 2019

  2. [2]

    Cox, M.L

    I.J. Cox, M.L. Miller, T.P. Minka, T.V. Papathomas, and P.N. Yianilos. 2000. The Bayesian image retrieval system, PicHunter: theory, implementation, and psy- chophysical experiments. IEEE Transactions on Image Processing 9, 1 (2000), 20–37. doi:10.1109/83.817596

    work page doi:10.1109/83.817596 2000

  3. [3]

    Cox, M.L

    I.J. Cox, M.L. Miller, T.P. Minka, and P.N. Yianilos. 1998. An optimized interaction strategy for Bayesian relevance feedback. In Proceedings. 1998 IEEE Computer So- ciety Conference on Computer Vision and Pattern Recognition (Cat. No.98CB36231) . 553–558. doi:10.1109/CVPR.1998.698660

    work page arXiv 1998

  4. [4]

    Cox, M.L

    I.J. Cox, M.L. Miller, S.M. Omohundro, and P.N. Yianilos. 1996. PicHunter: Bayesian relevance feedback for image retrieval. In Proceedings of 13th Inter- national Conference on Pattern Recognition , Vol. 3. 361–369 vol.3. doi:10.1109/ ICPR.1996.546971

    work page arXiv 1996

  5. [5]

    Zhijian Hou, Chong-Wah Ngo, and W. K. Chan. 2021. CONQUER: Contextual Query-aware Ranking for Video Corpus Moment Retrieval . Vol. 1. Association for Computing Machinery. 3900–3908 pages. doi:10.1145/3474085.3475281

    arXiv 2021

  6. [6]

    Miroslav Kratochvil, František Mejzlík, Patrik Veselý, Tomáš Soućek, and Jakub Lokoć. 2020. SOMHunter: Lightweight Video Search System with SOM-Guided Relevance Feedback. In Proceedings of the 28th ACM International Conference on Multimedia. 4481–4484. doi:10.1145/3394171.3414542

    work page arXiv 2020

  7. [7]

    Miroslav Kratochvíl, Patrik Veselý, František Mejzlík, and Jakub Lokoč. 2020. SOM-Hunter: Video Browsing with Relevance-to-SOM Feedback Loop. In Multi- Media Modeling: 26th International Conference, MMM 2020, Daejeon, South Korea, January 5–8, 2020, Proceedings, Part II . Springer-Verlag, 790–795. doi:10.1007/978- 3-030-37734-2_71

    doi:10.1007/978- 2020

  8. [8]

    Berg, and Mohit Bansal

    Jie Lei, Licheng Yu, Tamara L. Berg, and Mohit Bansal. 2020. TVR: A Large-Scale Dataset for Video-Subtitle Moment Retrieval. In Computer Vision – ECCV 2020: 16th European Conference, Glasgow, UK, August 23–28, 2020, Proceedings, Part XXI . Springer-Verlag, Berlin, Heidelberg, 447–463. doi:10.1007/978-3-030-58589-1_27

    work page doi:10.1007/978-3-030-58589-1_27 2020

Show all 34 references
  1. [9]

    Junnan Li, Dongxu Li, Silvio Savarese, and Steven Hoi. 2023. BLIP-2: bootstrap- ping language-image pre-training with frozen image encoders and large language models. In Proceedings of the 40th International Conference on Machine Learning (ICML’23). JMLR.org, Article 814, 13 pages

    2023

  2. [10]

    Junnan Li, Dongxu Li, Caiming Xiong, and Steven Hoi. 2022. BLIP: Bootstrap- ping Language-Image Pre-training for Unified Vision-Language Understanding and Generation. In Proceedings of the 39th International Conference on Machine Learning, Vol. 162. PMLR, 12888–12900

    2022

  3. [11]

    Linjie Li, Yen-Chun Chen, Yu Cheng, Zhe Gan, Licheng Yu, and Jingjing Liu

  4. [12]

    Jakub Lokoč, Stelios Andreadis, Werner Bailer, Aaron Duane, Cathal Gurrin, Zhixin Ma, Nicola Messina, Thao-Nhu Nguyen, Ladislav Peška, Luca Rossetto, Loris Sauter, Konstantin Schall, Klaus Schoeffmann, Omar Shahbaz Khan, Florian Spiess, Lucia Vadicamo, and Stefanos Vrochidis. ...

    2023 doi

  5. [13]

    Jakub Lokoč, Zuzana Vopálková, Patrik Dokoupil, and Ladislav Peška. 2023. Video Search with CLIP and Interactive Text Query Reformulation. In MultiMedia Mod- eling, Duc-Tien Dang-Nguyen, Cathal Gurrin, Martha Larson, Alan F. Smeaton, Stevan Rudinac, Minh-Son Dao, Christoph Tra...

    2023

  6. [14]

    Jakub Lokoč, František Mejzlík, Tomáš Souček, Patrik Dokoupil, and Ladislav Peška. 2022. Video Search with Context-Aware Ranker and Relevance Feedback. In MultiMedia Modeling: 28th International Conference, MMM 2022, Phu Quoc, Vietnam, June 6–10, 2022, Proceedings, Part II . S...

    2022 doi

  7. [15]

    Jakub Lokoč, Patrik Veselý, František Mejzlík, Gregor Kovalčík, Tomáš Souček, Luca Rossetto, Klaus Schoeffmann, Werner Bailer, Cathal Gurrin, Loris Sauter, Jaeyub Song, Stefanos Vrochidis, Jiaxin Wu, and Björn þóR Jónsson. 2021. Is the Reign of Interactive Search Eternal? Find...

    2021

  8. [16]

    Huaishao Luo, Lei Ji, Ming Zhong, Yang Chen, Wen Lei, Nan Duan, and Tianrui Li. 2022. CLIP4Clip: An empirical study of CLIP for end to end video clip retrieval and captioning. Neurocomput. 508, C (Oct. 2022), 293–304. doi:10.1016/j.neucom. 2022.07.028

    2022 doi

  9. [17]

    Multimedia Comput

    ACM Trans. Multimedia Comput. Commun. Appl. 17, 3, Article 91 (July 2021), 26 pages. doi:10.1145/3445031

    2021 doi

  10. [18]

    Luca Rossetto, Heiko Schuldt, George Awad, and Asad A. Butt. 2019. V3C – A Research Video Collection. In MultiMedia Modeling , Ioannis Kompatsiaris, Benoit Huet, Vasileios Mezaris, Cathal Gurrin, Wen-Huang Cheng, and Stefanos Vrochidis (Eds.). Springer International Publishing...

    2019

  11. [19]

    Luca Rossetto, Klaus Schoeffmann, and Abraham Bernstein. 2021. Insights on the V3C2 Dataset. CoRR abs/2105.01475 (2021). arXiv:2105.01475

    2021 arXiv

  12. [20]

    Smeulders, M

    A.W.M. Smeulders, M. Worring, S. Santini, A. Gupta, and R. Jain. 2000. Content- based image retrieval at the end of the early years. IEEE Transactions on Pattern Analysis and Machine Intelligence 22, 12 (2000), 1349–1380. doi:10.1109/34.895972

    2000 doi

  13. [21]

    Huang, M

    Yong Rui, T.S. Huang, M. Ortega, and S. Mehrotra. 1998. Relevance feedback: a power tool for interactive content-based image retrieval. IEEE Transactions on Circuits and Systems for Video Technology 8, 5 (1998), 644–655. doi:10.1109/76. 718510

    1998 doi

  14. [22]

    Lucia Vadicamo, Rahel Arnold, Werner Bailer, Fabio Carrara, Cathal Gurrin, Nico Hezel, Xinghan Li, Jakub Lokoc, Sebastian Lubos, Zhixin Ma, Nicola Messina, Thao-Nhu Nguyen, Ladislav Peska, Luca Rossetto, Loris Sauter, Klaus Schöffmann, Florian Spiess, Minh-Triet Tran, and Stef...

    2024

  15. [23]

    Li, and Shaoping Ma

    Zhong Su, Hongjiang Zhang, S. Li, and Shaoping Ma. 2003. Relevance feedback in content-based image retrieval: Bayesian framework, feature subspaces, and progressive learning. IEEE Transactions on Image Processing 12, 8 (2003), 924–937. doi:10.1109/TIP.2003.815254

    2003

  16. [24]

    Jiaxin Wu and Chong Wah Ngo. 2020. Interpretable Embedding for Ad-Hoc Video Search. MM 2020 - Proceedings of the 28th ACM International Conference on Multimedia (2020), 3357–3366. doi:10.1145/3394171.3413916

    2020

  17. [25]

    Patrik Veselý, František Mejzlík, and Jakub Lokoč. 2021. SOMHunter V2 at Video Browser Showdown 2021. In MultiMedia Modeling, Jakub Lokoč, Tomáš Skopal, Klaus Schoeffmann, Vasileios Mezaris, Xirong Li, Stefanos Vrochidis, and Ioannis Patras (Eds.). Springer International Publi...

    2021

  18. [26]

    Rintaro Yanagi, Ren Togo, Takahiro Ogawa, and Miki Haseyama. 2021. Interactive re-ranking for cross-modal retrieval based on object-wise question answering. In Proceedings of the 2nd ACM International Conference on Multimedia in Asia (Virtual Event, Singapore) (MMAsia ’20). As...

    2021

  19. [27]

    Rong Yan, Alexander Hauptmann, and Rong Jin. 2003. Multimedia search with pseudo-relevance feedback. In Proceedings of the 2nd International Conference on Image and Video Retrieval (CIVR’03). Springer-Verlag, Berlin, Heidelberg, 238–247

    2003

  20. [28]

    Huang, S

    Yong Rui Yong Rui, T.S. Huang, S. Mehrotra, and M. Ortega. 1997. A relevance feedback architecture for content-based multimedia information retrieval sys- tems. In 1997 Proceedings IEEE Workshop on Content-Based Access of Image and Video Libraries. 82–89. doi:10.1109/IVL.1997.629724

    1997

  21. [29]

    Kim-Hui Yap and Kui Wu. 2005. A soft relevance framework in content-based image retrieval systems. IEEE Transactions on Circuits and Systems for Video Technology 15, 12 (2005), 1557–1568. doi:10.1109/TCSVT.2005.856912

    2005

  22. [30]

    Xiang Sean Zhou and Thomas S. Huang. 2003. Relevance feedback in image retrieval: A comprehensive review. Multimedia Systems 8, 6 (01 Apr 2003), 536–

    2003

  23. [31]

    Yuanhan Zhang, Bo Li, haotian Liu, Yong jae Lee, Liangke Gui, Di Fu, Jiashi Feng, Ziwei Liu, and Chunyuan Li. 2024. LLaVA-NeXT: A Strong Zero-shot Video Understanding Model. https://llava-vl.github.io/blog/2024-04-30-llava-next- video

    2024

  24. [544]

    Ours”) is compared with three ablated ver- sions, each removing a specific component: “SoftUpd

    doi:10.1007/s00530-002-0070-3 ICMR ’25, June 30-July 3, 2025, Chicago, IL, USA Zhixin Ma and Chong-Wah Ngo A Ablation Study Tables 4 and 5 provide results from ablation studies that analyze the contribution of different components to the overall model per- formance. Table 4 re...

    2025 doi

  25. [2020]

    In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP)

    HERO: Hierarchical Encoder for Video+Language Omni-representation Pre-training. In Proceedings of the 2020 Conference on Empirical Methods in Natural Language Processing (EMNLP). 2046–2065. doi:10.18653/v1/2020.emnlp-main.161

    2020 doi

  26. [2024]

    Close-up of motorbike exhaust pipes being cleaned with a wet sponge. Two chromed pipes are visible, open on the left

    As discussed in Section 5.5, the full query is released in three rounds. We use colored solid, dashed, and dotted underlines to denote the query segments newly introduced in the first, second, and third rounds, respectively. Additionally, Figure 7 provides an example of user f...

    2025

Pith tools

Reviewed August 7, 2026 · model on record in the stance chip above.